Skip to Main Content

Recoup is the SLA-credit claims agent that remembers every win and loss in MongoDB, and whose next claim is better because of its last one. It reads GitHub's and Vercel's live public incident history, decides which incidents qualify for service credits under each vendor's published SLA, and has a deterministic checker do the uptime math, credit tier and filing deadline. The harness is the point: every decision, retrieved precedent, vendor response and reasoning step is stored in MongoDB Atlas (cases with embeddings, versioned playbooks, replayable traces); the agent retrieves similar past cases with Atlas Vector Search before deciding; and when a claim is rejected, a critic rewrites the playbook, with some rules becoming deterministic guardrails. Run the same incidents again and the agent claims more of what the vendor accepts and less of what it rejects, and the UI shows the lesson it wrote and the precedent it cited. Honesty notes: vendor outcomes are produced by a deterministic reviewer applying the real SLA text (real vendors may differ); spend figures are example data; all code was written today, Oct 11, 2026.

Team

nikhilk

More from The Harness Engineering & Model Wrangling #SFTechWeek Hackathon

MongoMation is an autonomous PLC engineer. You describe a machine in plain English, and Claude agents write IEC 61131-3 PLC code, compile and deploy it to a…

MongoMation project preview

LabRat is a self-improving harness for a physical lab agent, running on a real bench: two SO-101 arms, a 3018 CNC carrying a servo pipette I designed during…

LabRat project preview

Recursive Harness Scientist: an AI scientist that improves its own scientific method, and must replicate every gain on data it has never seen. Why it…

Recursive Harness Scientist project preview

Super Bid is a self-improving AI estimator that teaches itself to price bids the way a specific company would. It practices on the company’s real past bids…

SuperBid project preview

A robot that gets better at handling objects it’s never seen, without retraining. After every round, an AI coach reviews what broke and rewrites the robot’s…

WORLDSTATE gives physical spaces a memory. Any phone becomes an observer: open a link, point the camera, and WORLDSTATE tracks what's inside fridges, shelves…

WorldSTATE project preview